4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Portions Copyright 2011 Martin Matuska
24 * Copyright (c) 2012, 2017 by Delphix. All rights reserved.
27 #include <sys/zfs_context.h>
28 #include <sys/txg_impl.h>
29 #include <sys/dmu_impl.h>
30 #include <sys/dmu_tx.h>
31 #include <sys/dsl_pool.h>
32 #include <sys/dsl_scan.h>
34 #include <sys/callb.h>
37 * ZFS Transaction Groups
38 * ----------------------
40 * ZFS transaction groups are, as the name implies, groups of transactions
41 * that act on persistent state. ZFS asserts consistency at the granularity of
42 * these transaction groups. Each successive transaction group (txg) is
43 * assigned a 64-bit consecutive identifier. There are three active
44 * transaction group states: open, quiescing, or syncing. At any given time,
45 * there may be an active txg associated with each state; each active txg may
46 * either be processing, or blocked waiting to enter the next state. There may
47 * be up to three active txgs, and there is always a txg in the open state
48 * (though it may be blocked waiting to enter the quiescing state). In broad
49 * strokes, transactions -- operations that change in-memory structures -- are
50 * accepted into the txg in the open state, and are completed while the txg is
51 * in the open or quiescing states. The accumulated changes are written to
52 * disk in the syncing state.
56 * When a new txg becomes active, it first enters the open state. New
57 * transactions -- updates to in-memory structures -- are assigned to the
58 * currently open txg. There is always a txg in the open state so that ZFS can
59 * accept new changes (though the txg may refuse new changes if it has hit
60 * some limit). ZFS advances the open txg to the next state for a variety of
61 * reasons such as it hitting a time or size threshold, or the execution of an
62 * administrative action that must be completed in the syncing state.
66 * After a txg exits the open state, it enters the quiescing state. The
67 * quiescing state is intended to provide a buffer between accepting new
68 * transactions in the open state and writing them out to stable storage in
69 * the syncing state. While quiescing, transactions can continue their
70 * operation without delaying either of the other states. Typically, a txg is
71 * in the quiescing state very briefly since the operations are bounded by
72 * software latencies rather than, say, slower I/O latencies. After all
73 * transactions complete, the txg is ready to enter the next state.
77 * In the syncing state, the in-memory state built up during the open and (to
78 * a lesser degree) the quiescing states is written to stable storage. The
79 * process of writing out modified data can, in turn modify more data. For
80 * example when we write new blocks, we need to allocate space for them; those
81 * allocations modify metadata (space maps)... which themselves must be
82 * written to stable storage. During the sync state, ZFS iterates, writing out
83 * data until it converges and all in-memory changes have been written out.
84 * The first such pass is the largest as it encompasses all the modified user
85 * data (as opposed to filesystem metadata). Subsequent passes typically have
86 * far less data to write as they consist exclusively of filesystem metadata.
88 * To ensure convergence, after a certain number of passes ZFS begins
89 * overwriting locations on stable storage that had been allocated earlier in
90 * the syncing state (and subsequently freed). ZFS usually allocates new
91 * blocks to optimize for large, continuous, writes. For the syncing state to
92 * converge however it must complete a pass where no new blocks are allocated
93 * since each allocation requires a modification of persistent metadata.
94 * Further, to hasten convergence, after a prescribed number of passes, ZFS
95 * also defers frees, and stops compressing.
97 * In addition to writing out user data, we must also execute synctasks during
98 * the syncing context. A synctask is the mechanism by which some
99 * administrative activities work such as creating and destroying snapshots or
100 * datasets. Note that when a synctask is initiated it enters the open txg,
101 * and ZFS then pushes that txg as quickly as possible to completion of the
102 * syncing state in order to reduce the latency of the administrative
103 * activity. To complete the syncing state, ZFS writes out a new uberblock,
104 * the root of the tree of blocks that comprise all state stored on the ZFS
105 * pool. Finally, if there is a quiesced txg waiting, we signal that it can
106 * now transition to the syncing state.
109 static void txg_sync_thread(void *arg
);
110 static void txg_quiesce_thread(void *arg
);
112 int zfs_txg_timeout
= 5; /* max seconds worth of delta per txg */
115 * Prepare the txg subsystem.
118 txg_init(dsl_pool_t
*dp
, uint64_t txg
)
120 tx_state_t
*tx
= &dp
->dp_tx
;
122 bzero(tx
, sizeof (tx_state_t
));
124 tx
->tx_cpu
= kmem_zalloc(max_ncpus
* sizeof (tx_cpu_t
), KM_SLEEP
);
126 for (c
= 0; c
< max_ncpus
; c
++) {
129 mutex_init(&tx
->tx_cpu
[c
].tc_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
130 mutex_init(&tx
->tx_cpu
[c
].tc_open_lock
, NULL
, MUTEX_DEFAULT
,
132 for (i
= 0; i
< TXG_SIZE
; i
++) {
133 cv_init(&tx
->tx_cpu
[c
].tc_cv
[i
], NULL
, CV_DEFAULT
,
135 list_create(&tx
->tx_cpu
[c
].tc_callbacks
[i
],
136 sizeof (dmu_tx_callback_t
),
137 offsetof(dmu_tx_callback_t
, dcb_node
));
141 mutex_init(&tx
->tx_sync_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
143 cv_init(&tx
->tx_sync_more_cv
, NULL
, CV_DEFAULT
, NULL
);
144 cv_init(&tx
->tx_sync_done_cv
, NULL
, CV_DEFAULT
, NULL
);
145 cv_init(&tx
->tx_quiesce_more_cv
, NULL
, CV_DEFAULT
, NULL
);
146 cv_init(&tx
->tx_quiesce_done_cv
, NULL
, CV_DEFAULT
, NULL
);
147 cv_init(&tx
->tx_exit_cv
, NULL
, CV_DEFAULT
, NULL
);
149 tx
->tx_open_txg
= txg
;
153 * Close down the txg subsystem.
156 txg_fini(dsl_pool_t
*dp
)
158 tx_state_t
*tx
= &dp
->dp_tx
;
161 ASSERT0(tx
->tx_threads
);
163 mutex_destroy(&tx
->tx_sync_lock
);
165 cv_destroy(&tx
->tx_sync_more_cv
);
166 cv_destroy(&tx
->tx_sync_done_cv
);
167 cv_destroy(&tx
->tx_quiesce_more_cv
);
168 cv_destroy(&tx
->tx_quiesce_done_cv
);
169 cv_destroy(&tx
->tx_exit_cv
);
171 for (c
= 0; c
< max_ncpus
; c
++) {
174 mutex_destroy(&tx
->tx_cpu
[c
].tc_open_lock
);
175 mutex_destroy(&tx
->tx_cpu
[c
].tc_lock
);
176 for (i
= 0; i
< TXG_SIZE
; i
++) {
177 cv_destroy(&tx
->tx_cpu
[c
].tc_cv
[i
]);
178 list_destroy(&tx
->tx_cpu
[c
].tc_callbacks
[i
]);
182 if (tx
->tx_commit_cb_taskq
!= NULL
)
183 taskq_destroy(tx
->tx_commit_cb_taskq
);
185 kmem_free(tx
->tx_cpu
, max_ncpus
* sizeof (tx_cpu_t
));
187 bzero(tx
, sizeof (tx_state_t
));
191 * Start syncing transaction groups.
194 txg_sync_start(dsl_pool_t
*dp
)
196 tx_state_t
*tx
= &dp
->dp_tx
;
198 mutex_enter(&tx
->tx_sync_lock
);
200 dprintf("pool %p\n", dp
);
202 ASSERT0(tx
->tx_threads
);
206 tx
->tx_quiesce_thread
= thread_create(NULL
, 0, txg_quiesce_thread
,
207 dp
, 0, &p0
, TS_RUN
, minclsyspri
);
210 * The sync thread can need a larger-than-default stack size on
211 * 32-bit x86. This is due in part to nested pools and
212 * scrub_visitbp() recursion.
214 tx
->tx_sync_thread
= thread_create(NULL
, 32<<10, txg_sync_thread
,
215 dp
, 0, &p0
, TS_RUN
, minclsyspri
);
217 mutex_exit(&tx
->tx_sync_lock
);
221 txg_thread_enter(tx_state_t
*tx
, callb_cpr_t
*cpr
)
223 CALLB_CPR_INIT(cpr
, &tx
->tx_sync_lock
, callb_generic_cpr
, FTAG
);
224 mutex_enter(&tx
->tx_sync_lock
);
228 txg_thread_exit(tx_state_t
*tx
, callb_cpr_t
*cpr
, kthread_t
**tpp
)
230 ASSERT(*tpp
!= NULL
);
233 cv_broadcast(&tx
->tx_exit_cv
);
234 CALLB_CPR_EXIT(cpr
); /* drops &tx->tx_sync_lock */
239 txg_thread_wait(tx_state_t
*tx
, callb_cpr_t
*cpr
, kcondvar_t
*cv
, clock_t time
)
241 CALLB_CPR_SAFE_BEGIN(cpr
);
244 (void) cv_timedwait(cv
, &tx
->tx_sync_lock
,
245 ddi_get_lbolt() + time
);
247 cv_wait(cv
, &tx
->tx_sync_lock
);
249 CALLB_CPR_SAFE_END(cpr
, &tx
->tx_sync_lock
);
253 * Stop syncing transaction groups.
256 txg_sync_stop(dsl_pool_t
*dp
)
258 tx_state_t
*tx
= &dp
->dp_tx
;
260 dprintf("pool %p\n", dp
);
262 * Finish off any work in progress.
264 ASSERT3U(tx
->tx_threads
, ==, 2);
267 * We need to ensure that we've vacated the deferred space_maps.
269 txg_wait_synced(dp
, tx
->tx_open_txg
+ TXG_DEFER_SIZE
);
272 * Wake all sync threads and wait for them to die.
274 mutex_enter(&tx
->tx_sync_lock
);
276 ASSERT3U(tx
->tx_threads
, ==, 2);
280 cv_broadcast(&tx
->tx_quiesce_more_cv
);
281 cv_broadcast(&tx
->tx_quiesce_done_cv
);
282 cv_broadcast(&tx
->tx_sync_more_cv
);
284 while (tx
->tx_threads
!= 0)
285 cv_wait(&tx
->tx_exit_cv
, &tx
->tx_sync_lock
);
289 mutex_exit(&tx
->tx_sync_lock
);
293 txg_hold_open(dsl_pool_t
*dp
, txg_handle_t
*th
)
295 tx_state_t
*tx
= &dp
->dp_tx
;
296 tx_cpu_t
*tc
= &tx
->tx_cpu
[CPU_SEQID
];
299 mutex_enter(&tc
->tc_open_lock
);
300 txg
= tx
->tx_open_txg
;
302 mutex_enter(&tc
->tc_lock
);
303 tc
->tc_count
[txg
& TXG_MASK
]++;
304 mutex_exit(&tc
->tc_lock
);
313 txg_rele_to_quiesce(txg_handle_t
*th
)
315 tx_cpu_t
*tc
= th
->th_cpu
;
317 ASSERT(!MUTEX_HELD(&tc
->tc_lock
));
318 mutex_exit(&tc
->tc_open_lock
);
322 txg_register_callbacks(txg_handle_t
*th
, list_t
*tx_callbacks
)
324 tx_cpu_t
*tc
= th
->th_cpu
;
325 int g
= th
->th_txg
& TXG_MASK
;
327 mutex_enter(&tc
->tc_lock
);
328 list_move_tail(&tc
->tc_callbacks
[g
], tx_callbacks
);
329 mutex_exit(&tc
->tc_lock
);
333 txg_rele_to_sync(txg_handle_t
*th
)
335 tx_cpu_t
*tc
= th
->th_cpu
;
336 int g
= th
->th_txg
& TXG_MASK
;
338 mutex_enter(&tc
->tc_lock
);
339 ASSERT(tc
->tc_count
[g
] != 0);
340 if (--tc
->tc_count
[g
] == 0)
341 cv_broadcast(&tc
->tc_cv
[g
]);
342 mutex_exit(&tc
->tc_lock
);
344 th
->th_cpu
= NULL
; /* defensive */
348 * Blocks until all transactions in the group are committed.
350 * On return, the transaction group has reached a stable state in which it can
351 * then be passed off to the syncing context.
354 txg_quiesce(dsl_pool_t
*dp
, uint64_t txg
)
356 tx_state_t
*tx
= &dp
->dp_tx
;
357 int g
= txg
& TXG_MASK
;
361 * Grab all tc_open_locks so nobody else can get into this txg.
363 for (c
= 0; c
< max_ncpus
; c
++)
364 mutex_enter(&tx
->tx_cpu
[c
].tc_open_lock
);
366 ASSERT(txg
== tx
->tx_open_txg
);
368 tx
->tx_open_time
= gethrtime();
370 DTRACE_PROBE2(txg__quiescing
, dsl_pool_t
*, dp
, uint64_t, txg
);
371 DTRACE_PROBE2(txg__opened
, dsl_pool_t
*, dp
, uint64_t, tx
->tx_open_txg
);
374 * Now that we've incremented tx_open_txg, we can let threads
375 * enter the next transaction group.
377 for (c
= 0; c
< max_ncpus
; c
++)
378 mutex_exit(&tx
->tx_cpu
[c
].tc_open_lock
);
381 * Quiesce the transaction group by waiting for everyone to txg_exit().
383 for (c
= 0; c
< max_ncpus
; c
++) {
384 tx_cpu_t
*tc
= &tx
->tx_cpu
[c
];
385 mutex_enter(&tc
->tc_lock
);
386 while (tc
->tc_count
[g
] != 0)
387 cv_wait(&tc
->tc_cv
[g
], &tc
->tc_lock
);
388 mutex_exit(&tc
->tc_lock
);
393 txg_do_callbacks(list_t
*cb_list
)
395 dmu_tx_do_callbacks(cb_list
, 0);
397 list_destroy(cb_list
);
399 kmem_free(cb_list
, sizeof (list_t
));
403 * Dispatch the commit callbacks registered on this txg to worker threads.
405 * If no callbacks are registered for a given TXG, nothing happens.
406 * This function creates a taskq for the associated pool, if needed.
409 txg_dispatch_callbacks(dsl_pool_t
*dp
, uint64_t txg
)
412 tx_state_t
*tx
= &dp
->dp_tx
;
415 for (c
= 0; c
< max_ncpus
; c
++) {
416 tx_cpu_t
*tc
= &tx
->tx_cpu
[c
];
418 * No need to lock tx_cpu_t at this point, since this can
419 * only be called once a txg has been synced.
422 int g
= txg
& TXG_MASK
;
424 if (list_is_empty(&tc
->tc_callbacks
[g
]))
427 if (tx
->tx_commit_cb_taskq
== NULL
) {
429 * Commit callback taskq hasn't been created yet.
431 tx
->tx_commit_cb_taskq
= taskq_create("tx_commit_cb",
432 max_ncpus
, minclsyspri
, max_ncpus
, max_ncpus
* 2,
436 cb_list
= kmem_alloc(sizeof (list_t
), KM_SLEEP
);
437 list_create(cb_list
, sizeof (dmu_tx_callback_t
),
438 offsetof(dmu_tx_callback_t
, dcb_node
));
440 list_move_tail(cb_list
, &tc
->tc_callbacks
[g
]);
442 (void) taskq_dispatch(tx
->tx_commit_cb_taskq
, (task_func_t
*)
443 txg_do_callbacks
, cb_list
, TQ_SLEEP
);
448 txg_is_syncing(dsl_pool_t
*dp
)
450 tx_state_t
*tx
= &dp
->dp_tx
;
451 ASSERT(MUTEX_HELD(&tx
->tx_sync_lock
));
452 return (tx
->tx_syncing_txg
!= 0);
456 txg_is_quiescing(dsl_pool_t
*dp
)
458 tx_state_t
*tx
= &dp
->dp_tx
;
459 ASSERT(MUTEX_HELD(&tx
->tx_sync_lock
));
460 return (tx
->tx_quiescing_txg
!= 0);
464 txg_has_quiesced_to_sync(dsl_pool_t
*dp
)
466 tx_state_t
*tx
= &dp
->dp_tx
;
467 ASSERT(MUTEX_HELD(&tx
->tx_sync_lock
));
468 return (tx
->tx_quiesced_txg
!= 0);
472 txg_sync_thread(void *arg
)
474 dsl_pool_t
*dp
= arg
;
475 spa_t
*spa
= dp
->dp_spa
;
476 tx_state_t
*tx
= &dp
->dp_tx
;
478 uint64_t start
, delta
;
480 txg_thread_enter(tx
, &cpr
);
484 uint64_t timeout
= zfs_txg_timeout
* hz
;
487 uint64_t dirty_min_bytes
=
488 zfs_dirty_data_max
* zfs_dirty_data_sync_pct
/ 100;
491 * We sync when we're scanning, there's someone waiting
492 * on us, or the quiesce thread has handed off a txg to
493 * us, or we have reached our timeout.
495 timer
= (delta
>= timeout
? 0 : timeout
- delta
);
496 while (!dsl_scan_active(dp
->dp_scan
) &&
497 !tx
->tx_exiting
&& timer
> 0 &&
498 tx
->tx_synced_txg
>= tx
->tx_sync_txg_waiting
&&
499 !txg_has_quiesced_to_sync(dp
) &&
500 dp
->dp_dirty_total
< dirty_min_bytes
) {
501 dprintf("waiting; tx_synced=%llu waiting=%llu dp=%p\n",
502 tx
->tx_synced_txg
, tx
->tx_sync_txg_waiting
, dp
);
503 txg_thread_wait(tx
, &cpr
, &tx
->tx_sync_more_cv
, timer
);
504 delta
= ddi_get_lbolt() - start
;
505 timer
= (delta
> timeout
? 0 : timeout
- delta
);
509 * Wait until the quiesce thread hands off a txg to us,
510 * prompting it to do so if necessary.
512 while (!tx
->tx_exiting
&& !txg_has_quiesced_to_sync(dp
)) {
513 if (tx
->tx_quiesce_txg_waiting
< tx
->tx_open_txg
+1)
514 tx
->tx_quiesce_txg_waiting
= tx
->tx_open_txg
+1;
515 cv_broadcast(&tx
->tx_quiesce_more_cv
);
516 txg_thread_wait(tx
, &cpr
, &tx
->tx_quiesce_done_cv
, 0);
520 txg_thread_exit(tx
, &cpr
, &tx
->tx_sync_thread
);
523 * Consume the quiesced txg which has been handed off to
524 * us. This may cause the quiescing thread to now be
525 * able to quiesce another txg, so we must signal it.
527 ASSERT(tx
->tx_quiesced_txg
!= 0);
528 txg
= tx
->tx_quiesced_txg
;
529 tx
->tx_quiesced_txg
= 0;
530 tx
->tx_syncing_txg
= txg
;
531 DTRACE_PROBE2(txg__syncing
, dsl_pool_t
*, dp
, uint64_t, txg
);
532 cv_broadcast(&tx
->tx_quiesce_more_cv
);
534 dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
535 txg
, tx
->tx_quiesce_txg_waiting
, tx
->tx_sync_txg_waiting
);
536 mutex_exit(&tx
->tx_sync_lock
);
538 start
= ddi_get_lbolt();
540 delta
= ddi_get_lbolt() - start
;
542 mutex_enter(&tx
->tx_sync_lock
);
543 tx
->tx_synced_txg
= txg
;
544 tx
->tx_syncing_txg
= 0;
545 DTRACE_PROBE2(txg__synced
, dsl_pool_t
*, dp
, uint64_t, txg
);
546 cv_broadcast(&tx
->tx_sync_done_cv
);
549 * Dispatch commit callbacks to worker threads.
551 txg_dispatch_callbacks(dp
, txg
);
556 txg_quiesce_thread(void *arg
)
558 dsl_pool_t
*dp
= arg
;
559 tx_state_t
*tx
= &dp
->dp_tx
;
562 txg_thread_enter(tx
, &cpr
);
568 * We quiesce when there's someone waiting on us.
569 * However, we can only have one txg in "quiescing" or
570 * "quiesced, waiting to sync" state. So we wait until
571 * the "quiesced, waiting to sync" txg has been consumed
572 * by the sync thread.
574 while (!tx
->tx_exiting
&&
575 (tx
->tx_open_txg
>= tx
->tx_quiesce_txg_waiting
||
576 txg_has_quiesced_to_sync(dp
)))
577 txg_thread_wait(tx
, &cpr
, &tx
->tx_quiesce_more_cv
, 0);
580 txg_thread_exit(tx
, &cpr
, &tx
->tx_quiesce_thread
);
582 txg
= tx
->tx_open_txg
;
583 dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
584 txg
, tx
->tx_quiesce_txg_waiting
,
585 tx
->tx_sync_txg_waiting
);
586 tx
->tx_quiescing_txg
= txg
;
588 mutex_exit(&tx
->tx_sync_lock
);
589 txg_quiesce(dp
, txg
);
590 mutex_enter(&tx
->tx_sync_lock
);
593 * Hand this txg off to the sync thread.
595 dprintf("quiesce done, handing off txg %llu\n", txg
);
596 tx
->tx_quiescing_txg
= 0;
597 tx
->tx_quiesced_txg
= txg
;
598 DTRACE_PROBE2(txg__quiesced
, dsl_pool_t
*, dp
, uint64_t, txg
);
599 cv_broadcast(&tx
->tx_sync_more_cv
);
600 cv_broadcast(&tx
->tx_quiesce_done_cv
);
605 * Delay this thread by delay nanoseconds if we are still in the open
606 * transaction group and there is already a waiting txg quiescing or quiesced.
607 * Abort the delay if this txg stalls or enters the quiescing state.
610 txg_delay(dsl_pool_t
*dp
, uint64_t txg
, hrtime_t delay
, hrtime_t resolution
)
612 tx_state_t
*tx
= &dp
->dp_tx
;
613 hrtime_t start
= gethrtime();
615 /* don't delay if this txg could transition to quiescing immediately */
616 if (tx
->tx_open_txg
> txg
||
617 tx
->tx_syncing_txg
== txg
-1 || tx
->tx_synced_txg
== txg
-1)
620 mutex_enter(&tx
->tx_sync_lock
);
621 if (tx
->tx_open_txg
> txg
|| tx
->tx_synced_txg
== txg
-1) {
622 mutex_exit(&tx
->tx_sync_lock
);
626 while (gethrtime() - start
< delay
&&
627 tx
->tx_syncing_txg
< txg
-1 && !txg_stalled(dp
)) {
628 (void) cv_timedwait_hires(&tx
->tx_quiesce_more_cv
,
629 &tx
->tx_sync_lock
, delay
, resolution
, 0);
632 mutex_exit(&tx
->tx_sync_lock
);
636 txg_wait_synced(dsl_pool_t
*dp
, uint64_t txg
)
638 tx_state_t
*tx
= &dp
->dp_tx
;
640 ASSERT(!dsl_pool_config_held(dp
));
642 mutex_enter(&tx
->tx_sync_lock
);
643 ASSERT3U(tx
->tx_threads
, ==, 2);
645 txg
= tx
->tx_open_txg
+ TXG_DEFER_SIZE
;
646 if (tx
->tx_sync_txg_waiting
< txg
)
647 tx
->tx_sync_txg_waiting
= txg
;
648 dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
649 txg
, tx
->tx_quiesce_txg_waiting
, tx
->tx_sync_txg_waiting
);
650 while (tx
->tx_synced_txg
< txg
) {
651 dprintf("broadcasting sync more "
652 "tx_synced=%llu waiting=%llu dp=%p\n",
653 tx
->tx_synced_txg
, tx
->tx_sync_txg_waiting
, dp
);
654 cv_broadcast(&tx
->tx_sync_more_cv
);
655 cv_wait(&tx
->tx_sync_done_cv
, &tx
->tx_sync_lock
);
657 mutex_exit(&tx
->tx_sync_lock
);
661 txg_wait_open(dsl_pool_t
*dp
, uint64_t txg
)
663 tx_state_t
*tx
= &dp
->dp_tx
;
665 ASSERT(!dsl_pool_config_held(dp
));
667 mutex_enter(&tx
->tx_sync_lock
);
668 ASSERT3U(tx
->tx_threads
, ==, 2);
670 txg
= tx
->tx_open_txg
+ 1;
671 if (tx
->tx_quiesce_txg_waiting
< txg
)
672 tx
->tx_quiesce_txg_waiting
= txg
;
673 dprintf("txg=%llu quiesce_txg=%llu sync_txg=%llu\n",
674 txg
, tx
->tx_quiesce_txg_waiting
, tx
->tx_sync_txg_waiting
);
675 while (tx
->tx_open_txg
< txg
) {
676 cv_broadcast(&tx
->tx_quiesce_more_cv
);
677 cv_wait(&tx
->tx_quiesce_done_cv
, &tx
->tx_sync_lock
);
679 mutex_exit(&tx
->tx_sync_lock
);
683 * If there isn't a txg syncing or in the pipeline, push another txg through
684 * the pipeline by queiscing the open txg.
687 txg_kick(dsl_pool_t
*dp
)
689 tx_state_t
*tx
= &dp
->dp_tx
;
691 ASSERT(!dsl_pool_config_held(dp
));
693 mutex_enter(&tx
->tx_sync_lock
);
694 if (!txg_is_syncing(dp
) &&
695 !txg_is_quiescing(dp
) &&
696 tx
->tx_quiesce_txg_waiting
<= tx
->tx_open_txg
&&
697 tx
->tx_sync_txg_waiting
<= tx
->tx_synced_txg
&&
698 tx
->tx_quiesced_txg
<= tx
->tx_synced_txg
) {
699 tx
->tx_quiesce_txg_waiting
= tx
->tx_open_txg
+ 1;
700 cv_broadcast(&tx
->tx_quiesce_more_cv
);
702 mutex_exit(&tx
->tx_sync_lock
);
706 txg_stalled(dsl_pool_t
*dp
)
708 tx_state_t
*tx
= &dp
->dp_tx
;
709 return (tx
->tx_quiesce_txg_waiting
> tx
->tx_open_txg
);
713 txg_sync_waiting(dsl_pool_t
*dp
)
715 tx_state_t
*tx
= &dp
->dp_tx
;
717 return (tx
->tx_syncing_txg
<= tx
->tx_sync_txg_waiting
||
718 tx
->tx_quiesced_txg
!= 0);
722 * Verify that this txg is active (open, quiescing, syncing). Non-active
723 * txg's should not be manipulated.
726 txg_verify(spa_t
*spa
, uint64_t txg
)
728 dsl_pool_t
*dp
= spa_get_dsl(spa
);
729 if (txg
<= TXG_INITIAL
|| txg
== ZILTEST_TXG
)
731 ASSERT3U(txg
, <=, dp
->dp_tx
.tx_open_txg
);
732 ASSERT3U(txg
, >=, dp
->dp_tx
.tx_synced_txg
);
733 ASSERT3U(txg
, >=, dp
->dp_tx
.tx_open_txg
- TXG_CONCURRENT_STATES
);
737 * Per-txg object lists.
740 txg_list_create(txg_list_t
*tl
, spa_t
*spa
, size_t offset
)
744 mutex_init(&tl
->tl_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
746 tl
->tl_offset
= offset
;
749 for (t
= 0; t
< TXG_SIZE
; t
++)
750 tl
->tl_head
[t
] = NULL
;
754 txg_list_destroy(txg_list_t
*tl
)
758 for (t
= 0; t
< TXG_SIZE
; t
++)
759 ASSERT(txg_list_empty(tl
, t
));
761 mutex_destroy(&tl
->tl_lock
);
765 txg_list_empty(txg_list_t
*tl
, uint64_t txg
)
767 txg_verify(tl
->tl_spa
, txg
);
768 return (tl
->tl_head
[txg
& TXG_MASK
] == NULL
);
772 * Returns true if all txg lists are empty.
774 * Warning: this is inherently racy (an item could be added immediately
775 * after this function returns). We don't bother with the lock because
776 * it wouldn't change the semantics.
779 txg_all_lists_empty(txg_list_t
*tl
)
781 for (int i
= 0; i
< TXG_SIZE
; i
++) {
782 if (!txg_list_empty(tl
, i
)) {
790 * Add an entry to the list (unless it's already on the list).
791 * Returns B_TRUE if it was actually added.
794 txg_list_add(txg_list_t
*tl
, void *p
, uint64_t txg
)
796 int t
= txg
& TXG_MASK
;
797 txg_node_t
*tn
= (txg_node_t
*)((char *)p
+ tl
->tl_offset
);
800 txg_verify(tl
->tl_spa
, txg
);
801 mutex_enter(&tl
->tl_lock
);
802 add
= (tn
->tn_member
[t
] == 0);
804 tn
->tn_member
[t
] = 1;
805 tn
->tn_next
[t
] = tl
->tl_head
[t
];
808 mutex_exit(&tl
->tl_lock
);
814 * Add an entry to the end of the list, unless it's already on the list.
815 * (walks list to find end)
816 * Returns B_TRUE if it was actually added.
819 txg_list_add_tail(txg_list_t
*tl
, void *p
, uint64_t txg
)
821 int t
= txg
& TXG_MASK
;
822 txg_node_t
*tn
= (txg_node_t
*)((char *)p
+ tl
->tl_offset
);
825 txg_verify(tl
->tl_spa
, txg
);
826 mutex_enter(&tl
->tl_lock
);
827 add
= (tn
->tn_member
[t
] == 0);
831 for (tp
= &tl
->tl_head
[t
]; *tp
!= NULL
; tp
= &(*tp
)->tn_next
[t
])
834 tn
->tn_member
[t
] = 1;
835 tn
->tn_next
[t
] = NULL
;
838 mutex_exit(&tl
->tl_lock
);
844 * Remove the head of the list and return it.
847 txg_list_remove(txg_list_t
*tl
, uint64_t txg
)
849 int t
= txg
& TXG_MASK
;
853 txg_verify(tl
->tl_spa
, txg
);
854 mutex_enter(&tl
->tl_lock
);
855 if ((tn
= tl
->tl_head
[t
]) != NULL
) {
856 ASSERT(tn
->tn_member
[t
]);
857 ASSERT(tn
->tn_next
[t
] == NULL
|| tn
->tn_next
[t
]->tn_member
[t
]);
858 p
= (char *)tn
- tl
->tl_offset
;
859 tl
->tl_head
[t
] = tn
->tn_next
[t
];
860 tn
->tn_next
[t
] = NULL
;
861 tn
->tn_member
[t
] = 0;
863 mutex_exit(&tl
->tl_lock
);
869 * Remove a specific item from the list and return it.
872 txg_list_remove_this(txg_list_t
*tl
, void *p
, uint64_t txg
)
874 int t
= txg
& TXG_MASK
;
875 txg_node_t
*tn
, **tp
;
877 txg_verify(tl
->tl_spa
, txg
);
878 mutex_enter(&tl
->tl_lock
);
880 for (tp
= &tl
->tl_head
[t
]; (tn
= *tp
) != NULL
; tp
= &tn
->tn_next
[t
]) {
881 if ((char *)tn
- tl
->tl_offset
== p
) {
882 *tp
= tn
->tn_next
[t
];
883 tn
->tn_next
[t
] = NULL
;
884 tn
->tn_member
[t
] = 0;
885 mutex_exit(&tl
->tl_lock
);
890 mutex_exit(&tl
->tl_lock
);
896 txg_list_member(txg_list_t
*tl
, void *p
, uint64_t txg
)
898 int t
= txg
& TXG_MASK
;
899 txg_node_t
*tn
= (txg_node_t
*)((char *)p
+ tl
->tl_offset
);
901 txg_verify(tl
->tl_spa
, txg
);
902 return (tn
->tn_member
[t
] != 0);
906 * Walk a txg list -- only safe if you know it's not changing.
909 txg_list_head(txg_list_t
*tl
, uint64_t txg
)
911 int t
= txg
& TXG_MASK
;
912 txg_node_t
*tn
= tl
->tl_head
[t
];
914 txg_verify(tl
->tl_spa
, txg
);
915 return (tn
== NULL
? NULL
: (char *)tn
- tl
->tl_offset
);
919 txg_list_next(txg_list_t
*tl
, void *p
, uint64_t txg
)
921 int t
= txg
& TXG_MASK
;
922 txg_node_t
*tn
= (txg_node_t
*)((char *)p
+ tl
->tl_offset
);
924 txg_verify(tl
->tl_spa
, txg
);
927 return (tn
== NULL
? NULL
: (char *)tn
- tl
->tl_offset
);